Active Nerf Eval

active-nerf-eval

ActiveNeRF: Learning Accurate 3D Geometry by Active Pattern Projection — Jianyu Tao et al. (2024) (arXiv:2408.06592, 2024)

What this evaluates

Evaluates the accuracy of 3D geometry reconstruction from multi-view images using active pattern projection. It measures how closely the predicted point cloud matches the ground truth geometry and assesses robustness under varying view counts and camera-projector baselines.

Datasets

• NeRF derivative (synthetic) — total 800; splits: test (800)
• Real-world capture (RealSense D415) — total 120; splits: test (120)

Metrics

• Chamfer Distance (mm) (primary) — range: other
  • Symmetric Chamfer Distance: CD = 1/(2N) Σ_i min_j ||p_i - q_j|| + 1/(2M) Σ_j min_i ||q_j - p_i||, where p and q are predicted and ground truth points. Reported in millimeters.
• P(CD < 0.01m) (%) — range: percent
  • Percentage of points in the predicted point cloud where the Chamfer Distance to the ground truth is less than 0.01 meters (10 mm).
• P(CD < 0.05m) (%) — range: percent
  • Percentage of points in the predicted point cloud where the Chamfer Distance to the ground truth is less than 0.05 meters (50 mm).

Input / output format

Input: Multi-view images (with and without active light projection), along with known camera and active light projector intrinsics and extrinsics. Images are downsampled to half resolution for input.

Output: Reconstructed 3D point cloud representing the scene geometry.

Scoring recipe

def evaluate(pred_pc, gt_pc, voxel_size=0.003):
    pred_pc = downsample_voxel(pred_pc, voxel_size)
    gt_pc = downsample_voxel(gt_pc, voxel_size)
    cd_m = chamfer_distance(pred_pc, gt_pc)
    cd_mm = cd_m * 1000
    p_01 = (cd_m < 0.01).mean() * 100
    p_05 = (cd_m < 0.05).mean() * 100
    return cd_mm, p_01, p_05

Common pitfalls

• Point clouds must be downsampled to a voxel size of 0.003 before computing Chamfer Distance to ensure resolution consistency.
• The baseline between the camera and active light projector is critical; a zero baseline (collocated) causes the method to fail completely.
• Non-Lambertian surfaces (e.g., specular) can degrade active light pattern learning unless a BRDF module is included.
• Input images are downsampled to half resolution during training, which may affect high-frequency detail recovery.

Evidence (verbatim from paper)

We use chamfer distance between our reconstruction and ground truth to evaluate our model performance. For both reconstruction and ground truth, we downsample the point clouds to a voxel size of 0.003 to ensure resolution consistency. To measure the reconstruction quality in more depth, we compute the percentage of chamfer distance that is lower than 0.01m and 0.05m.

Citation

@misc{tao2024activerf,
  title={ActiveNeRF: Learning Accurate 3D Geometry by Active Pattern Projection},
  author={Jianyu Tao et al. (2024)},
  year={2024},
  note={arXiv:2408.06592}
}

• arXiv: 2408.06592